ANCHORED GROUNDING RECEIPTS

Anchored Grounding Receipts

A signed, anchored receipt that the claims in a generated answer were bound to specific spans of named source documents, with a reproducible binding-integrity check and a method-attributed support score. Each citation hash-commits to the exact source span it points at, so anyone holding the sources can re-verify that cited claims resolve to real text. It records citation-binding provenance, never a guarantee that the answer is true.

AGrR · Shipped1 wire typeApache-2.0 · CC0 spec
← All familiesVerify a receipt →

WIRE TYPES

ar.grounding.v1 (grounding.rag)

Single canonical wire type for RAG answers. Commits to the full answer (SHA-256), lists each named source by id and document hash, each claim by id and commitment, and each citation with its source span offsets, a SHA-256 commit of that span, a support level (supported / partial / unsupported) and a score from a named scorer method. Carries deterministic coverage (fraction of claims cited) and grounded_count, scored at a UTC time.

WHAT IT PROVES

  • — Every cited claim points at a real, hash-committed span of a named source document — the binding is deterministic and re-checkable by anyone holding the sources (span_commit + source_commit).
  • — The recorded citations, claims, sources, and scores are exactly as the issuer sealed them and have not been altered (Ed25519 signature check).
  • — The receipt was signed by a registered issuer key (Ed25519 signature check).
  • — The coverage metric — the fraction of answer claims carrying at least one resolved citation — is arithmetically derivable from the committed body (deterministic count over committed fields).

WHAT IT DOESN'T PROVE

  • — That the answer is true, or that the cited sources are themselves accurate — binding a claim to a span is not verifying the claim.
  • — That an uncited or unsupported claim is false — coverage measures citation, not correctness.
  • — That the support score reflects real semantic entailment — v1 scores lexical overlap by a named method; the score is provenance of measurement, advisory only.
  • — That the producer segmented the answer into claims correctly — the claim list is the producer's assertion, recorded verbatim by commitment, not adjudicated.

COMPOSES WITH

AGrR receipts reference other family members via body-level composition pointers — verifier-coordinated, not signature-mandated.

AQRAnchored Quality Receipts

Both record output-side provenance for a generated answer: AQR scores it against an eval suite, AGrR binds its claims to sources — quality provenance and grounding provenance for the same output.

AERAnchored Evaluation Receipts

Deliberately distinct: AER carries EU AI Act accuracy and transparency attestations at the system level; AGrR carries per-answer citation binding. Neither family carries the other's claims.

Verify any
AGrR receipt.

verify.dekimu.com ↗

Paste any claim ID to verify a receipt, check its anchor, and inspect the issuer signature.

REFERENCES

AI Act Art. 13 — Transparency and provision of information to deployers (EUR-Lex) ↗
AI Act Art. 15 — Accuracy, robustness and cybersecurity (EUR-Lex) ↗

Anchored Grounding Receipts are cryptographic provenance and privacy-lifecycle protocols. verify.dekimu.com is a reference implementation, not a qualified trust service under Regulation (EU) No 910/2014 (eIDAS) or successor.